Testul de scalare a legilor pentru rețelele neuronale

Răspunsurile și explicațiile apar pe măsură ce susțineți testul.

1. What pattern did Kaplan et al. report for language-model loss as model size, data, or compute increased?
   • Approximate power-law relationships over the tested ranges
   • A sudden guaranteed drop to zero
   • A fixed increase in loss for every larger model
   • Random oscillation with no measurable trend
2. Care trei mărimi se referă legile de scalare inițiale la pierdere?
   • Mărimea lotului, abandonul și epocile
   • Parametrii modelului, dimensiunea setului de date și calculul
   • Rata de învățare, scăderea greutății și adâncimea
   • Număr de GPU, memorie și latență
3. How can a scaling-law fit help when planning a larger training run?
   • It guarantees the final benchmark score
   • It eliminates the need for validation
   • It can inform an extrapolation from smaller experiments, subject to the tested setup
   • It predicts results exactly for any dataset or architecture
4. Ce reprezintă termenul constant (ireductibil) dintr-o formulă de pierdere a legii de scalare?
   • O eroare de software
   • Rata de învățare
   • Numărul de GPU
   • Entropia intrinsecă a datelor pe care niciun model nu o poate învinge
5. Why can absolute loss reductions become smaller as compute keeps increasing under a power-law fit?
   • The reducible loss term shrinks while the loss approaches its floor
   • The data is removed after each compute increase
   • The model stops using parameters
   • Power laws require loss to increase over time
6. Care tendință împinge domeniul dincolo de simpla scalare „mai mare este mai bine”?
   • GPU-urile devin libere
   • Lipsa datelor de înaltă calitate plus limite de energie și costuri
   • Apare text nelimitat de înaltă calitate
   • Curbe de pierdere care se transformă în linii verticale drepte
7. What did the original Kaplan study say about architecture details such as width and depth within its tested range?
   • They always dominated model size, data, and compute
   • They made loss independent of data
   • They had relatively small effects compared with the studied scaling variables
   • They guaranteed the same performance for all architectures
8. What compute-allocation tendency did the original Kaplan paper report?
   • Every model should be trained to zero loss
   • Small models always use less compute at equal loss
   • Training data has no effect on compute allocation
   • Large models could be more sample-efficient and compute-efficient training could stop before convergence